1. science
2. mathematics
3. statistics

Sample Path Properties of Bifractional Brownian Motion
Ciprian A. Tudor
SAMOS-MATISSE, Centre d’Economie de La Sorbonne,
Universit´e de Panth´eon-Sorbonne Paris 1,
90, rue de Tolbiac, 75634 Paris Cedex 13, France.
[email protected]
URL: http://pageperso.aol.fr/cipriantudor/
Yimin Xiao ∗
Department of Statistics and Probability,
A-413 Wells Hall, Michigan State University,
East Lansing, MI 48824, USA.
[email protected]
URL: http://www.stt.msu.edu/~xiaoyimi
June 29, 2006
©
ª
Abstract
Let B H,K = B H,K (t), t ∈ R+ be a bifractional Brownian motion in Rd . We prove
that B H,K is strongly locally nondeterministic. Applying this property and a stochastic
integral representation of B H,K , we establish Chung’s law of the iterated logarithm for
B H,K , as well as sharp H¨older conditions and tail probability estimates for the local times
of B H,K .
We also consider the existence and the
© regularity of the
ª localdtimes of multiparameter
bifractional Brownian motion B H,K = B H,K (t), t ∈ RN
in R using Wiener-Itˆo chaos
+
expansion.
Running head: Sample Path Properties of Bifractional Brownian Motion
2000 AMS Classification Numbers: Primary 60G15, 60G17.
Key words: Bifractional Brownian motion, self-similar Gaussian processes, small ball
probability, Chung’s law of the iterated logarithm, local times, level set, Hausdorff dimension,
chaos expansion, multiple Wiener-Itˆo stochastic integrals.
∗
Research partially supported by the NSF grant DMS-0404729.
1
1
Introduction
In recent years, there has been of considerable interest in studying fractional Brownian motion
due to its applications in various scientific areas including telecommunications, turbulence,
image processing and finance. On the other hand, many authors have proposed to use more
general self-similar Gaussian processes and random fields as stochastic models; see e.g. Addie
et al. (1999), Anh et al. (1999), Benassi et al. (2000), Mannersalo and Norros (2002), Bonami
and Estrade (2003), Cheridito (2004), Benson et al. (2006). Such applications have raised
many interesting theoretical questions about self-similar Gaussian processes and fields in
general. However, contrast to the extensive studies on fractional Brownian motion, there has
been little systematic investigation on other self-similar Gaussian processes. The main reasons
for this, in our opinion, are the complexity of dependence structures and the non-availability
of convenient stochastic integral representations for self-similar Gaussian processes which do
not have stationary increments.
The objective of this paper is to fill this gap by developing systematic ways to study sample
path properties of self-similar Gaussian processes. Our main tools are the Lamperti transformation [which provides a powerful connection between self-similar processes and stationary
processes; see Lamperti (1962)] and the strong local nondeterminism of Gaussian processes
[see Xiao (2005)]. In particular, for any self-similar Gaussian process X = {X(t), t ∈ R}, the
Lamperti transformation leads to a stochastic integral representation for X. We will show
the usefulness of such a representation in studying sample path properties of X.
For concreteness, we only consider a rather special class of self-similar Gaussian processes,
namely, the bifractional Brownian motions introduced by Houdr´e and Villa (2003), to illustrate our methods. Given constants H ∈ (0, 1) and K ∈ (0, 1], the bifractional Brownian
motion (bi-fBm, in short) in R is a centered Gaussian process B0H,K = {B0H,K (t), t ∈ R+ }
with covariance function
i
1 h
RH,K (s, t) := R(s, t) = K (t2H + s2H )K − |t − s|2HK
(1.1)
2
and B0H,K (0) = 0.
Let B©1H,K , . . . , BdH,K be
independent copies of B0H,K . We define the Gaussian process
ª
B H,K = B H,K (t), t ∈ R+ with values in Rd by
³
´
B H,K (t) = B1H,K (t), . . . , BdH,K (t) ,
∀t ∈ R+ .
(1.2)
By (1.1) one can verify easily that B H,K is a self-similar process with index HK, that is, for
every constant a > 0,
o
n
o n
d
(1.3)
B H,K (at), t ∈ R+ = aHK B H,K (t), t ∈ R+ ,
d
where X = Y means the two processes have the same finite dimensional distributions. Note
that, when K = 1, B H,K is the ordinary fractional Brownian motion in Rd . However, if
K 6= 1, B H,K does not have stationary increments. In fact, fractional Brownian motion is the
2
only Gaussian self-similar process with stationary increments [see Samorodnitsky and Taqqu
(1994)].
Russo and Tudor (2006) have established some properties on the strong variations, local
times and stochastic calculus of real-valued bifractional Brownian motion. An interesting
property that deserves to be recalled is the fact that, when HK = 12 , the quadratic variation
of this process on [0, t] is equal to a constant times t. This is really remarkable since as far as
we know this is the only Gaussian self-similar process with this quadratic variation besides
Brownian motion. Taking into account this property, it is natural to ask if the bifractional
Brownian motion B H,K with KH = 21 shares other properties with Brownian motion (from
the sample path regularity point of view). As it can be seen from the rest of the paper, the
answer is often positive: for example, the bi-fBm with HK = 21 and Brownian motion satisfy
the same forms of Chung’s laws of the iterated logarithm and the H¨older conditions for their
local times.
The rest of this paper is organized as follows. In Section 2 we apply the Lamperti
transformation to prove the strong local nondeterminism of B0H,K . This property will play
essential roles in proving most of our results. In Section 3 we derive small ball probability
estimates and a stochastic integral representation for B0H,k . Applying these results, we prove
a Chung’s law of the iterated logarithm for bifractional Brownian motion.
Section 4 is devoted to the study of local times of one-parameter bifractional Brownian
motion and the corresponding N -parameter fields. In general, there are mainly two methods
in studying local times of Gaussian processes: the Fourier analysis approach introduced by
Berman and the Malliavin calculus approach. It is known that, the Fourier analysis approach
combined with various properties of local nondeterminism yields strong regularity properties
such as the joint continuity and sharp H¨older conditions for the local times [see Berman
(1973), Pitt (1978), Geman and Horowitz (1980), Xiao (1997, 2005)]; while the Malliavin
calculus approach requires less conditions on the process and establishes regularity of the
local times in the sense of Sobolev-Watanabe spaces [see Watanabe (1984), Imkeller et al.
(1995), Eddahbi et al. (2005)]. In this paper we make use of both approaches to obtain more
comprehensive results on local times of bifractional Brownian motion and fields.
Throughout this paper, an unspecified positive and finite constant will be denoted by
c, which may not be the same in each occurrence. More specific constants in Section i are
numbered as ci,1 , ci,2 , . . ..
2
Strong local nondeterminism
The following proposition is essential in this paper. From its proof, we see that the same
conclusion holds for quite general self-similar Gaussian processes.
Proposition 2.1 For all constants 0 < a < b, B0H,K is strongly locally ϕ-nondeterministic
on I = [a, b] with ϕ(r) = r2HK . That is, there exist positive constants c2,1 and r0 such that
for all t ∈ I and all 0 < r ≤ min{t, r0 },
³
´
¯
Var B0H,K (t)¯B0H,K (s) : s ∈ I, r ≤ |s − t| ≤ r0 ≥ c2,1 ϕ(r).
(2.1)
3
Proof We consider the centered stationary Gaussian process Y0 = {Y0 (t), t ∈ R} defined
through the Lamperti’s transformation [Lamperti (1962)]:
Y0 (t) = e−HK t B0H,K (et ), ∀ t ∈ R.
¡
¢
The covariance function r(t) := E Y0 (0)Y0 (t) is given by
·
¸
¯2HK
¢K ¯ t
1 −HKt ¡ 2Ht
¯
¯
r(t) = K e
e
+1 − e −1
2
·
¸
¯
¯
¢
1 HKt ¡
−2Ht K
−t ¯2HK
¯
1+e
− 1−e
= Ke
.
2
(2.2)
(2.3)
Hence r(t) is an even function and, by (2.3) and the Taylor expansion, we verify that r(t) =
O(e−βt ) as t → ∞, where β = min{H(2 − K), HK}. It follows that r(·) ∈ L1 (R). On the
other hand, by using (2.3) and the Taylor expansion again, we also have
r(t) ∼ 1 −
1
|t|2HK
2K
as t → 0.
(2.4)
The stationary Gaussian process Y0 is sometimes called the Ornstein-Uhlenbeck process
associated with B0H,K [Note that it does not coincide with the solution of the fractional
Langevin equation, see Cheridito et al. (2003) for a proof in the case K = 1]. By Bochner’s
theorem, Y0 has the following stochastic integral representation
Z
Y0 (t) =
eiλt W (dλ),
∀ t ∈ R,
(2.5)
R
where W is a complex Gaussian measure with control measure ∆ whose Fourier transform is
r(·). The measure ∆ is called the spectral measure of Y .
Since r(·) ∈ L1 (R), so the spectral measure ∆ of Y has a continuous density function
f (λ) which can be represented as the inverse Fourier transform of r(·):
Z
1 ∞
f (λ) =
r(t) cos(tλ) dt.
(2.6)
π 0
It follows from (2.4), (2.6) and the Tauberian theorem due to Pitman (1968, Theorem 5) [cf.
Bingham et al. (1987)] that
f (λ) ∼ c2,2 |λ|−(1+2HK)
as λ → ∞,
(2.7)
where c2,2 > 0 is an explicit constant depending only on HK. Hence, by a result of Cuzick and
DuPreez (1982, Lemma 1) [see also Xiao (2005) for more general results], Y0 = {Y0 (t), t ∈ R}
is strongly locally ϕ-nondeterministic on any interval J = [−T, T ] with ϕ(r) = r2HK in
the sense that there exist positive constants δ and c2,3 such that for all t ∈ [−T, T ] and all
r ∈ (0, |t| ∧ δ),
¯
¡
¢
Var Y0 (t)¯Y0 (s) : s ∈ J, r ≤ |s − t| ≤ δ ≥ c2,3 ϕ(r).
(2.8)
4
Now we prove the strong local nondeterminism of B0H,K on I. To this end, note that
B0H,K (t) = tHK Y0 (log t) for all t > 0. We choose r0 = aδ. Then for all s, t ∈ I with
r ≤ |s − t| ≤ r0 we have
¯
r ¯¯
≤ log s − log t¯ ≤ δ.
(2.9)
b
Hence it follows from (2.8) and (2.9) that for all t ∈ [a, b] and r < r0 ,
³
´
¯
Var B0H,K (t)¯B0H,K (s) : s ∈ I, r ≤ |s − t| ≤ r0
³
´
¯
= Var tHK Y0 (log t)¯sHK Y0 (log s) : s ∈ I, r ≤ |s − t| ≤ r0
´
³
¯
(2.10)
≥ t2HK Var Y0 (log t)¯Y0 (log s) : s ∈ I, r ≤ |s − t| ≤ r0
³
´
¯
≥ a2HK Var Y0 (log t)¯Y0 (log s) : s ∈ I, r/b ≤ | log s − log t| ≤ δ
≥ c2,4 ϕ(r).
This proves Proposition 2.1.
¤
For use in next section, we list two properties of the spectral density f (λ) of Y . They
follow from (2.7) or, more generally, from (2.4) and the truncation inequalities in Lo´eve (1977,
p.209); see also Monrad and Rootz´en (1995).
Lemma 2.2 There exist positive constants c2,5 and c2,6 such that for u > 1,
Z
λ2 f (λ) dλ ≤ c2,5 u2(1−HK)
(2.11)
|λ|<u
and
Z
|λ|≥u
f (λ) dλ ≤ c2,6 u−2HK .
(2.12)
We will also need the following lemma from Houdr´e and Villa (2003).
Lemma 2.3 There exist positive constants c2,7 and c2,8 such that for all s, t ∈ R+ , we have
h³
´2 i
c2,7 |t − s|2HK ≤ E B0H,K (t) − B0H,K (s)
≤ c2,8 |t − s|2HK .
3
(2.13)
Chung’s law of the iterated logarithm
As applications of small ball probability estimates, Monrad and Rootzen (1995), Xiao (1997)
and Li and Shao (2001) established Chung-type laws of the iterated logarithm for fractional
Brownian motion and other strongly locally nondeterministic Gaussian processes with stationary increments. However, there have been no results on Chung’s LIL for self-similar Gaussian
processes that do not have stationary increments [Recall that the class of self-similar Gaussian
processes is large and fBm is the only such process with stationary increments].
5
In this section, we prove the following Chung’s law of the iterated logarithm for bifractional Brownian motion in R. It will be clear that our argument is applicable to a large class
of self-similar Gaussian processes.
Theorem 3.1 Let B0H,K = {B0H,K (t), t ∈ R+ } be a bifractional Brownian motion in R.
Then there exists a positive and finite constant c3,1 such that
¯
¯
maxt∈[0,r] ¯B0H,K (t)¯
lim inf HK
= c3,1
a.s.
(3.1)
r→0 r
/(log log(1/r))HK
In order to prove Theorem 3.1, we need several preliminary results. Lemma 3.2 gives
estimates on the small ball probability of B0H,K .
Lemma 3.2 There exist positive constants c3,2 and c3,3 such that for all t0 ∈ [0, 1] and
x ∈ (0, 1),
¾
µ
µ
¶
½
¶
¯ H,K
¯
c3,3
c3,2
H,K
¯
¯
exp − 1/(HK) ≤ P max B0 (t) − B0 (t0 ) ≤ x ≤ exp − 1/(HK) .
(3.2)
t∈[0,1]
x
x
Proof By Proposition 2.1 and Lemma 2.3, we see that B0H,K satisfies Conditions (C1) and
(C2) in Xiao (2005). Hence this lemma follows from Theorem 3.1 in Xiao (2005).
¤
Proposition 3.3 provides a zero-one law for ergodic self-similar processes, which complements the results of Takashima (1989). In order to state it, we need to recall some definitions.
Let X = {X(t), t ∈ R} be a separable, self-similar process with index κ. For any constant
a > 0, the scaling transformation Sκ,a of X is defined by
(Sκ,a X)(t) = a−κ X(at),
∀t ∈ R.
(3.3)
Note that X is κ-self-similar is equivalent to saying that for every a > 0, the process
{(Sκ,a X)(t), t ∈ R} has the same finite dimensional distributions as those of X. That is,
for a κ-self-similar process X, a scaling transformation Sκ,a preserves the distribution of
X, and so the notion of ergodicity and mixing of Sκ,a can be defined in the usual way, cf.
Cornfeld et al. (1982). Following Takashima (1989), we say that a κ-self-similar process
X = {X(t), t ∈ R} is ergodic (or strong mixing) if for every a > 0, a 6= 1, the scaling transformation Sκ,a is ergodic (or strong mixing, respectively). This, in turn, is equivalent to saying
that the shift transformations for the corresponding stationary process Y = {Y (t), t ∈ R}
defined by Y (t) = e−κt X(et ) are ergodic (or strong mixing, respectively).
Proposition 3.3 Let X = {X(t), t ∈ R} be a separable, self-similar process with index κ.
We assume that X(0) = 0 and X is ergodic. Then for any increasing function ψ : R+ → R+ ,
we have P(Eκ,ψ ) = 0 or 1, where
½
¾
κ
Eκ,ψ = ω : there exists δ > 0 such that sup |X(s)| ≥ t ψ(t) for all 0 < t ≤ δ . (3.4)
0≤s≤t
6
Proof We will prove that for every a > 0, the event Eκ,ψ is invariant with respective to
the transformation Sκ,a . Then the conclusion follows from the ergodicity of X.
Fix a constant a > 0 and a 6= 1. We consider two cases: (i) a > 1 and (ii) a < 1. In the
first case, since ψ is increasing, we have ψ(au) ≥ ψ(u) for all u > 0. Assume that a.s. there
is a δ > 0 such that
sup |X(s)| ≥ tκ ψ(t)
for all 0 < t ≤ δ,
(3.5)
0≤s≤t
then
¯
¯
sup ¯a−κ X(as)¯ = a−κ sup |X(s)| ≥ tκ ψ(t)
0≤s≤t
for all 0 < t ≤ δ/a.
(3.6)
0≤s≤at
¡
¢
−1 E
This implies that Eκ,ψ ⊂ Sκ,a
of X, these two events have
κ,ψ . ©By the self-similarity
¡
¢ª
−1 E
the same probability, it follows that P Eκ,ψ ∆Sκ,a
=
0. This proves that Eκ,ψ is
κ,ψ
Sκ,a -invariant and, hence, has probability 0 or 1.
In case (ii), we have ψ(au) ≤ ψ(u) for all u > 0 and the¡ proof
¢ is similar to the above. If
−1 E
Sκ,a X ∈ Eκ,ψ , then we have X ∈ Eκ,ψ . This implies Sκ,a
κ,ψ ⊂ Eκ,ψ and again Eκ,ψ is
Sκ,a -invariant. This finishes the proof.
¤
By a result of Manuyama (1949) on ergodicity and mixing properties of stationary Gaussian
processes, we see that B0H,K is mixing. Hence we have the following corollary of Proposition
3.3.
Corollary 3.4 There exists a constant c3,4 ∈ [0, ∞] such that
lim inf
t→0+
¯ H,K ¯
(log log 1/t)HK
¯B
max
(s)¯ = c3,4 ,
0
0≤s≤t
tHK
a.s.
(3.7)
¡
¢−HK
ª
ª
©
©
Proof We take ψc (t) = c log log 1/t
and define c3,4 = sup c ≥ 0 : P Eκ, ψc = 1 .
It can be verified that (3.7) follows from Proposition 3.3.
¤
It follows from Corollary 3.4 that Theorem 3.1 will be established if we show c3,4 ∈ (0, ∞).
This is where Lemma 3.2 and the following lemma from Talagrand (1995) are needed.
Lemma 3.5 Let X = {X(t), t ∈ R} be a centered Gaussian process in R and let S ⊂ R be a
closed set equipped with the canonical metric defined by
h ¡
¢2 i1/2
d(s, t) = E X(s) − X(t)
.
Then there exists a positive constants c3,5 such that for all u > 0,
(
µ
¶)
µ
¶
Z Dp
u2
P sup |X(s) − X(t)| ≥ c3,5 u +
log Nd (S, ε) dε
≤ exp − 2 ,
D
s, t∈S
0
(3.8)
where Nd (S, ε) denotes the smallest number of open d-balls of radius ε needed to cover S and
where D = sup{d(s, t) : s, t ∈ S} is the diameter of S.
Now we proceed to prove Theorem 3.1.
7
Proof of Theorem 3.1 We prove the lower bound first. For any integer n ≥ 1, let rn = e−n .
Let 0 < γ < c3,3 be a constant and consider the event
½
¾
¯ H,K ¯
HK HK
HK
¯
¯
An =
max B0 (s) ≤ γ
rn /(log log 1/rn )
.
0≤s≤rn
Then the self-similarity of B0H,K and Lemma 3.2 imply that
µ
¶
c3,3
P {An } ≤ exp −
log n = n−c3,3 /γ .
γ
P
Since nn=1 P {An } < ∞, the Borel-Cantelli lemma implies
¯
¯
maxs∈[0,rn ] ¯B0H,K (s)¯
≥ c3,3
a.s.
lim inf HK
n→∞ rn /(log log(1/rn ))HK
It follows from (3.10) and a standard monotonicity argument that
¯
¯
maxt∈[0,r] ¯B0H,K (t)¯
≥ c3,6
a.s.
lim inf HK
r→0 r
/(log log(1/r))HK
(3.9)
(3.10)
(3.11)
The upper bound is a little more difficult to prove due to the dependence structure of
B0H,K . In order to create independence, we will make use of the following stochastic integral
representation of B0H,K : for every t > 0,
Z
B0H,K (t) = tHK
eiλ log t W (dλ).
(3.12)
R
This follows from the spectral representation (2.5) of Y and its connection with B0H,K .
For every integer n ≥ 1, we take
tn = n−n
and
dn = nβ ,
(3.13)
where β > 0 is a constant whose value will be determined later. It is sufficient to prove that
there exists a finite constant c3,7 such that
¯
¯
maxs∈[0,tn ] ¯B0H,K (s)¯
lim inf HK
≤ c3,7
a.s.
(3.14)
n→∞ tn /(log log(1/tn ))HK
en by
Let us define two Gaussian processes Xn and X
Z
Xn (t) = tHK
eiλ log t W (dλ)
(3.15)
|λ|∈(dn−1 ,dn ]
and
Z
en (t) = tHK
X
eiλ log t W (dλ),
(3.16)
|λ|∈(d
/ n−1 ,dn ]
en (t) for all t ≥ 0. It is important to note that the
respectively. Clearly B0H,K (t) = Xn (t) + X
Gaussian processes Xn (n = 1, 2, . . .) are independent and, moreover, for every n ≥ 1, Xn
en are independent as well.
and X
¡
¢−HK
Denote h(r) = rHK log log 1/r
. We make the following two claims:
8
(i). There is a constant γ > 0 such that
½
¾
∞
X
¯
¯
HK
¯
¯
P max Xn (s) ≤ γ
h(tn ) = ∞.
n=1
(ii). For every ε > 0,
s∈[0,tn ]
½
¾
∞
X
¯
¯
en (s)¯ > ε h(tn ) < ∞.
P max ¯X
n=1
s∈[0,tn ]
(3.17)
(3.18)
Since the events in (3.17) are independent, we see that (3.14) follows from (3.17), (3.18) and
a standard Borel-Cantelli argument.
It remains to verify the claims (i) and (ii) above. By Lemma 3.2 and Anderson’s inequality
[see Anderson (1955)], we have
¾
¾
½
½
¯ H,K ¯
¯
¯
HK
HK
¯
¯
¯
¯
h(tn )
P max Xn (s) ≤ γ
h(tn ) ≥ P max B0 (s) ≤ γ
s∈[0,tn ]
s∈[0,tn ]
³ c
´
(3.19)
≥ exp − 3,2 log(n log n)
γ
¡
¢−c /γ
= n log n 3,2 .
Hence (i) holds for γ ≥ c3,2 .
In order to prove (ii), we divide [0, tn ] into pn + 1 non-overlapping subintervals Jn,j =
en on each of Jn,j . Let β > 0
[an,j−1 , an,j ], (i = 0, 1, . . . , pn ) and then apply Lemma 3.5 to X
−β
be the constant in (3.13) and we take Jn,0 = [0, tn n ]. After Jn,j has been defined, we
take an,j+1 = an,j (1 + n−β ). It can be verified that the number of such subintervals of [0, tn ]
satisfies the following bound:
pn + 1 ≤ c nβ log n.
(3.20)
Moreover, for every j ≥ 1, if s, t ∈ Jn,j and s < t, then we have t/s − 1 ≤ n−β and this yields
³t´
t − s ≤ s n−β and log
≤ n−β .
(3.21)
s
en satisfies
Lemma 2.3 implies that the canonical metric d for the process X
d(s, t) ≤ c |s − t|HK
for all s, t > 0
(3.22)
and d(0, s) ≤ c tHK
n−βHK for every s ∈ Jn,0 . It follows that D0 := sup{d(s, t); s, t ∈ Jn,0 } ≤
n
HK
−βHK
c tn n
and
tn n−β
(3.23)
Nd (Jn,0 , ε) ≤ c 1/(HK) .
ε
Some simple calculation yields
r
Z D0 q
Z tHK
³ t n−β ´
n−βHK
n
n
log Nd (Jn,0 , ε) dε ≤
log 1/(HK) dε
ε
0
0
Z 1r ³ ´
(3.24)
1
= tHK
n−βHK
log
du
n
u
0
= c3,8 tHK
n−βHK .
n
9
It follows from Lemma 3.5 and (3.24) that
¾
µ
½
¯
¯
e
¯
¯
P max Xn (s) > ε h(tn ) ≤ exp − c ¡
s∈Jn,0
¶
¢2HK .
log(n log n)
n2βHK
(3.25)
For every 1 ≤ j ≤ pn , we estimate the d-diameter of Jn,j . It follows from (3.16) that for
any s, t ∈ Jn,j with s < t,
¯
¯2
³
´2 Z
¯ HK iλ log t
¯
e
e
E Xn (s) − Xn (t) =
e
− sHK eiλ log s ¯ f (λ) dλ
¯t
|λ|≤dn−1
Z
¯
¯2
¯ HK iλ log t
(3.26)
HK iλ log s ¯
+
e
−s
e
¯t
¯ f (λ) dλ
|λ|>dn
:= I1 + I2 .
The second term is easy to estimate: for all s, t ∈ Jn,j ,
Z
2HK
I2 ≤ 4 tn
f (λ) dλ ≤ c3,9 t2HK
n−2βHK ,
n
(3.27)
|λ|>dn
where the last inequality follows from (2.12).
For the first term, we use the elementary inequality 1 − cos x ≤ x2 to derive that for all
s, t ∈ Jn,j with s < t,
·
¸
Z
³
¡ HK
¢
¡
t ¢´
HK 2
HK HK
I1 =
t
−s
f (λ) dλ
+ 2t
s
1 − cos λ log
s
|λ|≤dn−1
´2HK Z
³t
³t´ Z
(3.28)
−1
≤ s2HK
f (λ) dλ + 2t2HK log2
λ2 f (λ) dλ
s
s
R
|λ|≤dn−1
≤ c3,10 t2HK
n−2βHK ,
n
where, in deriving the last inequality, we have used (3.21) and (2.11), respectively.
It follows from (3.26), (3.27) and (3.28) that the d-diameter of Jn,j satisfies
Dj ≤ c3,11 tHK
n−βHK .
n
(3.29)
Hence, similar to (3.25), we use Lemma 3.5 and (3.29) to derive
½
¶
¾
µ
¯
¯
n2βHK
e
¯
¯
P max Xn (s) > ε h(tn ) ≤ exp − c ¡
¢2HK .
s∈Jn,j
log(n log n)
(3.30)
By combining (3.20), (3.25) and (3.30) we derive that for every ε > 0,
pn
∞
∞ X
n
o X
n
o
X
¯
¯
¯
¯
en (s)¯ > ε h(tn ) ≤
en (s)¯ > ε h(tn )
P max ¯X
P max ¯X
n=1
s∈[0,tn ]
n=1 j=0
≤c
∞
X
s∈Jn,j
µ
β
n log n exp
n=1
−c¡
n2βHK
¢2HK
log(n log n)
¶
(3.31)
< ∞.
This proves (3.18) and hence the theorem.
¤
10
Remark 3.6 Let t0 ∈ [0, 1] be fixed and we consider the process X = {X(t), t ∈ R+ } defined
by X(t) = B0H,K (t + t0 ) − B0H,K (t0 ). By applying Lemma 3.2 and modifying the proof of
Theorem 3.1, one can show that
¯
¯
maxt∈[0,r] ¯B0H,K (t + t0 ) − B0H,K (t0 )¯
−1
c3,12 ≤ lim inf
≤ c3,12
a.s.,
(3.32)
r→0
rHK /(log log(1/r))HK
where c3,12 > 1 is a constant depending on HK only.
Corresponding to Lemma 3.2, we can also consider the small ball probability of B0H,K
under the H¨older-type norm. For α ∈ (0, 1) and any function y ∈ C0 ([0, 1]), we consider the
α-H¨older norm of y defined by,
kykα =
sup
s,t∈[0,1],s6=t
|y(s) − y(t)|
.
|s − t|α
(3.33)
The following proposition extends the results of Stolz (1996) and Theorem 2.1 of Kuelbs,
Li and Shao (1995) to bifractional Brownian motion.
Proposition 3.7 Let B0H,K be a bifractional Brownian motion in R and α ∈ (0, HK). There
exist positive constants c3,13 and c3,14 such that for all ε ∈ (0, 1),
³
´
n
o
³
´
exp − c3,13 ε−1/(HK−α) ≤ P kB0H,K kα ≤ ε ≤ exp − c3,14 ε−1/(HK−α) .
(3.34)
Proof
4
It follows from Theorem 3.4 in Xiao (2005).
¤
Local times of bifractional Brownian motion
This section is devoted to the study of the local times of the bi-fBm both in the one-parameter
and multi-parameter cases. As we pointed out in the Introduction there are essentially two
ways to prove the existence and regularity properties of local times for Gaussian processes:
the first is related to the Fourier analysis and the local nondeterminism property; the second is
based on the Malliavin calculus and Wiener-Itˆo chaos expansion. We will apply the Fourier
analysis approach for the one-parameter case and the Malliavin calculus approach for the
multiparameter case.
4.1
The one-parameter case
Let B H,K = {B H,K (t), t ∈ R+ } be a bifractional Brownian motion with indices H and K in
Rd . For any closed interval I ⊂ R+ and for any x ∈ Rd , the local time L(x, I) of B H,K is
defined as the density of the occupation measure µI defined by
Z
¡
¢
µI (A) = 11A B H,K (s) ds,
A ∈ B(Rd ).
I
11
It can be shown [cf. Geman and Horowitz (1980) Theorem 6.4] that the following occupation
density formula hods: for every Borel function g(t, x) ≥ 0 on I × Rd ,
Z
Z Z
¡
¢
H,K
g t, B
(t) dt =
g(t, x)L(x, dt) dx.
(4.1)
Rd
I
I
Lemma 2.3 and Theorem 21.9 in Geman and Horowitz (1980) imply that if 1/(HK) > d
then B H,K has a local time L(x, t) := L(x, [0, t]), where (x, t) ∈ Rd × [0, ∞). In fact, more
regularity properties of L(x, t) can be derived from Theorem 3.14 in Xiao (2005) which we
summarize in the following theorem. Besides interest in their own right, such results are also
useful in studying the fractal properties of the sample paths of B H,K .
Theorem 4.1 Let B H,K = {B H,K (t), t ∈ R} be a bifractional Brownian motion with indices
H and K in Rd . If 1/(HK) > d, then the following properties hold:
(i) B H,K has a local time L(x, t) that is jointly continuous in (x, t) almost surely.
(ii) [Local H¨older condition] For every B ∈ B(R), let L∗ (B) = supx∈Rd L(x, B) be the
maximum local time. Then there exists a positive constant c4,1 such that for all t0 ∈ R+ ,
lim sup
r→0
L∗ (B(t0 , r))
≤ c4,1
ϕ1 (r)
a.s.
(4.2)
Here and in the sequel, B(t, r) = (t − r, t + r) and ϕ1 (r) = r1−HKd (log log 1/r)HKd .
(iii) [Uniform H¨older condition] For every finite interval I ⊆ R, there exists a positive finite
constant c4,2 such that
lim sup sup
r→0
t0 ∈I
L∗ (B(t0 , r))
≤ c4,2
ϕ2 (r)
a.s.,
(4.3)
where ϕ2 (r) = r1−HKd (log 1/r)HKd .
Proof By Proposition 2.1 and Lemma 2.3, we see that the conditions of Theorem 3.14 in
Xiao (2005) are satisfied. Hence the results follow.
¤
The following states that the local H¨older condition for the maximum local time is sharp.
Remark 4.2 By the definition of local times, we have that for every interval Q ⊆ R+ ,
Z
³
¯
¯´d
|Q| =
L(x, Q) dx ≤ L∗ (Q) · max ¯B H,K (s) − B H,K (t)¯ .
(4.4)
s,t∈Q
B H,K (Q)
By taking Q = B(t0 , r) in (4.4) and using (3.32) in Remark 3.6, we derive the lower bound
in the following
L∗ (B(t0 , r))
c4,3 ≤ lim sup
≤ c4,4
a.s.,
(4.5)
ϕ1 (r)
r→0
12
where c4,3 > 0 is a constant independent of t0 and the upper bound is given by (4.2). A
similar lower bound for (4.3) could also be established by using (4.4), if one proves that for
every interval I ⊆ R+ ,
lim inf inf
r→0
t∈I
|B H,K (s) − B H,K (t)|
≤ c4,5
rHk /(log 1/r)HK
s∈B(t,r)
max
a.s.
(4.6)
Theorem 4.1 can be applied to determine the Hausdorff dimension and Hausdorff measure
of the level set Zx = {t ∈ R+ : B H,K (t) = x}, where x ∈ Rd . See Berman (1972), Monrad and
Pitt (1987) and Xiao (1997, 2005). In the following theorem we prove a uniform Hausdorff
dimension result for the level sets of B H,K .
Theorem 4.3 If 1/(HK) > d, then with probability one,
dimH Zx = 1 − HKd
for all x ∈ Rd ,
(4.7)
where dimH denotes Hausdorff dimension.
Proof
It follows from Theorem 3.19 in Xiao (2005) that with probability one,
dimH Zx = 1 − HKd for all x ∈ O,
(4.8)
where O is the random open set defined by
o
[ n
O=
x ∈ Rd : L(x, [s, t]) > 0 .
s,t∈Q; s<t
Hence it only remains to show O = Rd a.s. For this purpose, we consider the stationary
Gaussian process Y = {Y (t), t ∈ R} defined by Y (t) = e−HKt B H,K (et ), using the Lamperti
transformation.
Note that the component processes of Y are independent and, as shown in the proof of
Proposition 2.1, they are strongly locally ϕ-nondeterministic with ϕ(r) = r2HK . It follows
from Theorem 3.14 in Xiao (2005) that Y has a jointly continuous local time LY (x, t), where
(x, t) ∈ Rd × R. From the proof of Proposition 2.1, it can be verified that Y satisfies the
conditions of Theorem 2 in Monrad and Pitt (1987), it follows that almost surely for every
y ∈ Rd , there exists a finite interval J ⊂ R such that LY (y, J) > 0.
On the other hand, by using the occupation density formula (4.1), we can verify that the
local times of B H,K and Y are related by the following equation: for all x ∈ Rd and finite
interval I = [a, b] ⊂ [0, ∞),
Z
L(x, I) =
e(1−HK)s LY (e−HKs x, ds).
(4.9)
[log a, log b]
Hence, there exists a.s. a finite interval I such that L(0, I) > 0. The continuity of L(x, I)
implies the a.s. existence of δ > 0 such that L(y, I) > 0 for all y ∈ Rd with |y| ≤ δ.
Observe that the scaling property of B H,K implies that for all constants c > 0, the scaled
local time c−(1−HKd) L(x, ct) is a version of L(c−HK x, t). It follows that a.s. for every x ∈ Rd ,
L(x, J) > 0 for some finite interval J ⊂ [0, ∞).
¤
13
Since there is little knowledge on the explicit distribution of L(0, 1), it is of interest in
estimating the tail probability P{L(0, 1) > x} as x → ∞. This problem has been considered
by Kasahara et al. (1999) for certain fractional Brownian motion and by Xiao (2005) for a
large class of Gaussian processes. Our next result is a consequence of Theorem 3.20 in Xiao
(2005).
Theorem 4.4 Let B H,K = {B H,K (t), t ∈ R} be a bifractional Brownian motion in Rd with
indices H and K. If 1/(HK) > d, then for x > 0 large enough,
©
ª
− log P L(0, 1) > x ³ xHK ,
(4.10)
where a(x) ³ b(x) means a(x)/b(x) is bounded from below and above for x large enough.
Proof By Proposition 2.1 and Lemma 2.3, we see that the conditions of Theorem 3.20 in
Xiao (2005) are satisfied. This proves (4.10).
¤
Let us also note that the existence of the jointly continuous version of the local time and
the self-similarity allow us to prove the following renormalization result. The case d = 1 has
been proved in Russo and Tudor (2006).
Proposition 4.5 If 1/(HK) > d, then for any integrable function F : Rd → R,
Z
¡
¢
(d)
HK−1
t
F B H,K (u) du −→ Fe L(0, 1) as t → ∞,
(4.11)
[0,t]
where Fe =
Proof
Z
R
Rd
F (x) dx.
It holds that
Z
¡
¢
F B H,K (u) du = t
[0,t]
¡
¢
d
F B H,K (tv) dv = t
[0,1]
Z
¡
¢
F tHK B H,K (v) dv.
By using the occupation density formula, we derive
Z
Z
Z
¡
¡
¢
F B H,K (u)) du = t
F tHK x L(x, 1) dx = t1−HK
[0,t]
(4.12)
[0,1]
Rd
F (y) L(yt−HK , 1) dy. (4.13)
Rd
Since y 7→ L(y, 1) is almost surely continuous and bounded, the dominated convergence
theorem implies that, as t → ∞, the last integral in (4.13) tends to Fe L(0, 1) almost surely.
This and (4.12) yield (4.11).
¤
4.2
Oscillation of bifractional Brownian motion
The oscillations of certain classes of stochastic processes, especially Gaussian processes, in
the measure space ([0, 1], λ1 ), where λ1 is the Lebesgue measure in R, have been studied,
among others, by Wschebor (1992) and Aza¨ıs and Wschebor (1996). The following is an
analogous result for bifractional Brownian motion.
14
Proposition 4.6 Let B H,K be a bi-fBm in R with indices H ∈ (0, 1) and K ∈ (0, 1]. For
every t ∈ [0, 1], let
B H,K (t + ε) − B H,K (t)
Zε (t) =
.
εHK
Then the following statements hold:
(i) For every integer k ≥ 1, almost surely,
Z 1
¡
¢k
Zε (t) dt → E(ρk )
as ε → 0,
0
where ρ is a centered normal random variable with variance σ 2 = 21−K .
(ii) For every interval J ⊂ [0, 1], almost surely, for every x ∈ R
λ1 {t ∈ J : Zε (t) ≤ x} → λ1 (J) P(ρ ≤ x)
Proof
Let us denote
Z
Y
ε,k
1
=
0
It is sufficient to prove that
¡
¢
Var Y ε,k ≤ c(k) εβ
as ε → 0.
(Zε (t))k dt.
for some c(k) and β > 0.
(4.14)
Then the conclusions (i) and (ii) will follow as in Aza¨ıs and Wschebor (1996) by the means
of a Borel-Cantelli argument.
Note that
Z 1Z 1
¡
¢
¡
¢
Cov Zε (u)k , Zε (u)k dudv.
Var Y ε,k =
0
0
We will make use of the fact that for a centered Gaussian vector (U, V ),
X
£
¤p £
¤k−p
¡
¢
c(p, k) Cov(U, V ) Var(U )Var(V )
Cov U k , V k =
.
1≤p≤k
Since the random variable Zε has clearly bounded variance [cf. Lemma 2.3], it suffices to
show that for every 1 ≤ p ≤ k,
Z 1Z 1
£ ¡
¢¤p
E Zε (u)Zε (v) dudv ≤ c4,6 εβ
(4.15)
0
We can write
Z 1Z 1
0
0
0
£ ¡
¢¤p
E Zε (u)Zε (v) dvdu = 2
Z
1Z u
£ ¡
¢¤p
1l(u−v<ε) E Zε (u)Zε (v) dvdu
0
0
Z 1Z u
£ ¡
¢¤p
+2
1l(u−v≥ε) E Zε (u)Zε (v) dvdu
0
:= A + B.
15
0
Clearly A ≤ c ε, hence it suffices to bound the term B. Note that
Z u Z v
¡
¢
1
∂2R
E Zε (u)Zε (v) = 2HK
dbda.
ε
u−ε v−ε ∂a∂b
Since
i
¢K−2 2H−1 2H−1
∂2R
2HK h¡ 2H
(a, b) = K
a + b2H
a
b
− (2HK − 1)|a − b|2HK−2 ,
∂a∂b
2
we have
¸p
Z u Z v
Z 1 Z u−ε ·
¡ 2H
¢
1
2H K−2 2H−1 2H−1
B ≤ c(p, H, K)
a +b
a
b
dbda dvdu
ε2HK u−ε v−ε
0
0
¸p
Z u Z v
Z 1 Z u−ε ·
1
2HK−2
|a − b|
dbda dvdu
+ c(p, H, K)
ε2HK u−ε v−ε
0
0
:= B1 + B2 .
The term B2 can be treated as in the fBm case [see Aza¨ıs and Wschebor (1996), Proposition
2.1] and we get B2 ≤ c εβ for some constant β > 0. Finally, since a2HK + b2HK ≥ aHK bHK ,
we can write
¸p
Z 1 Z u−ε ·
Z u Z v
1
HK−1 HK−1
B1 ≤ c(p, H, K)
a
b
dbda
dvdu
ε2HK u−ε v−ε
0
0
¶p µ HK
¶p
Z 1 Z u−ε µ HK
v
− (v − ε)HK
u
− (u − ε)HK
dvdu
= c(p, H, K)
εHK
εHK
0
0
"Z µ
#2
¶p
1
uHK − (u − ε)HK
≤c
dvdu .
εHK
0
A change of variable shows that B1 ≤ c ε2(1−HK) . Combining the above yields (4.15). Therefore, we have proved (4.14), and the proposition.
¤
The above result can be extended to obtain the almost sure weak approximation of the
occupation measure of the bi-fBm B H,K by means of normalized number of crossing of BεH,K ,
where
BεH,K represents the convolution of B H,K with an approximation of the identity Φε (t) =
¡
¢
1
t
ε Φ ε with Φ = 1l[−1,0] . If g is a real function defined on an interval I, then the number of
crossing of level u is
Nu (g, I) = #{t ∈ I, g(t) = u},
where #E denotes the cardinality of E.
Proposition 4.7 Almost surely for every continuous function f and for every bounded interval I ⊂ R+ ,
Z ∞
Z ∞
³ π ´1/2
ε1−HK
f (u)Nu (BεH,K , I) du →
f (u)L(u, I) du
as ε → 0.
2
−∞
−∞
16
Proof The arguments in Aza¨ıs and Wschebor (1996), Section 5, apply. Details are left to
the reader.
¤
4.3
The multi-parameter case
For any given vectors H = (H1 , . . . , HN ) ∈ (0, 1)N and K = (K1 , . . . , KN ) ∈ (0, 1]N , an
(N, d)-bifractional Brownian sheet B H,K = {B H,K (t), t ∈ RN
+ } is a centered Gaussian random
field in Rd with i.i.d. components whose covariance functions are given by
·
¸
N
³
´ Y
´
1 ³ 2Hj
2Hj Kj
2Hj Kj
E B1H,K (s)B1H,K (t) =
s
+
t
−
|t
−
s
|
.
j
j
j
j
Kj
2
j=1
(4.16)
It follows from (4.16) that, similar to an (N, d)-fractional Brownian sheet [cf. Xiao and Zhang
(2002), Ayache and Xiao (2005)], B H,K is operator-self-similar. However, it does not have
convenient stochastic integral representations which have played essential rˆoles in the studies
of fractional Brownian sheets. Nevertheless, we will prove that the sample path properties
of B H,K are very similar to those of fractional Brownian sheets, and we can describe the
anisotropic properties of B H,K in terms of the vectors H and K.
We start with the following useful lemma.
Lemma 4.8 For any ε > 0, there exist positive and finite constants c4,7 and c4,8 such that
for all s, t ∈ [ε, 1]N ,
c4,7
N
X
2Hj Kj
|sj − tj |
·³
N
´2 ¸
X
H,K
H,K
≤ c4,8
|sj − tj |2Hj Kj ,
≤ E B1 (s) − B1 (t)
(4.17)
j=1
j=1
and
c4,7
N
X
N
³
´
X
|sj − tj |2Hj Kj ≤ detCov B1H,K (s), B1H,K (t) ≤ c4,8
|sj − tj |2Hj Kj .
j=1
(4.18)
j=1
Here and in the sequel, detCov denotes determinant of the covariance matrix.
Proof We will make use of the following easily verifiable fact: For any Gaussian random
vector (Z1 , Z2 ),
detCov(Z1 , Z2 ) = Var(Z1 )Var(Z2 |Z1 ),
(4.19)
where Var(Z1 ) and Var(Z2 |Z1 ) denote the variance of Z1 and the conditional variance of Z2 ,
given Z1 , respectively.
By (4.19) we see that for all s, t ∈ [ε, 1]N ,
³
´
h
i
³
´
¯
detCov B1H,K (s), B1H,K (t) = E B1H,K (s)2 Var B1H,K (t)¯B1H,K (s)
h
i ·³
´2 ¸
(4.20)
H,K
H,K
H,K
2
≤ E B1 (s) E B1 (s) − B1 (t)
.
17
¡
¢
Since Var B1H,K (s) is bounded from above and below by positive and finite constants, it is
sufficient to prove the upper bound in (4.17) and the lower bound in (4.18).
When N = 1, Lemma 2.3, Proposition 2.1 and (4.19) imply that both (4.17) and (4.18)
hold. Next we show that, if the lemma holds for any B H,K with at most n parameters, then
it holds for B H,K with n + 1 parameters.
We verify the upper bound in (4.17) first. For any s, t ∈ [ε, 1]n+1 , let s0 = (s1 , . . . , sn , tn+1 ).
Then we have
·³
·³
´2 ¸
´2 ¸
H,K
H,K
H,K
H,K 0
≤ 2E B1 (s) − B1 (s )
E B1 (s) − B1 (t)
·³
(4.21)
´2 ¸
H,K 0
H,K
+ 2E B1 (s ) − B1 (t)
.
For the first term, we note that whenever s1 , . . . , sn ∈ [ε, 1] are fixed, B H,K is a (rescaled)
bifractional Brownian motion in sn=1 . Hence Lemma 2.3 implies the first term in the righthand side of (4.21) is bounded by c |tn − sn |2Hn+1 Kn+1 , where the constant c is independent
of s1 , . . . , sn ∈ [ε, 1]. On the other hand, when tn+1 ∈ [ε, 1] is fixed, B H,K is a (rescaled)
(N, d)-bifractional Brownian sheet. Hence the induction
hypothesis implies the second term
P
in the right-hand side of (4.21) is bounded by c nj=1 |tj − sj |2Hj Kj . This and (4.21) together
prove the upper bound in (4.17).
Suppose the lower bound¡ in (4.18) holds for¢ any B H,K with at most n parameters. For
N = n + 1, we write detCov B1H,K (s), B1H,K (t) as
n+1
Y
2Hj Kj
tj
2Hj Kj
sj
−
j=1
=
n+1
Y
j=1
n+1
Y
½
2Hj Kj 2Hj Kj
1 K1 2H1 K1
tj
sj
s2H
t1
−
1
j=2
+
×
1
·
22K1
½ n+1
Y
¡ 2H1
¢K
t1 + s12H1 1
2H K
tj j j
2H K
sj j j
j=2
≥c
i2
1 h¡ 2Hj
2Hj ¢Kj
2Hj Kj
t
+
s
−
|t
−
s
|
j
j
j
j
22Kj
n+1
X
−
i2 ¾
1 h¡ 2H1
2H1 ¢K1
2H1 K1
t
+ s1
− |t1 − s1 |
22K1 1
¸2
2H1 K1
− |t1 − s1 |
n+1
Y
j=2
(4.22)
i2 ¾
1 h¡ 2Hj
2Hj ¢Kj
2Hj Kj
tj + sj
− |tj − sj |
22Kj
|sj − tj |2Hj Kj
j=1
where the last inequality follows from the induction hypothesis. This proves the lower bound
in (4.18).
¤
Applying Lemma 4.8, we can prove that many results in Xiao and Zhang (2002), Ayache
and Xiao (2005) on sample path properties of fractional Brownian sheet also hold for B H,K
as well. Theorem 4.9 is concerned with the existence of local times of B H,K .
18
©
ª
Theorem 4.9 Let B H,K = B H,K (t), t ∈ RN
+ be an (N, d)-bifractional Brownian sheet with
P
1
parameters H ∈ (0, 1)N and K ∈ (0, 1]N . If d < N
j=1 Hj Kj then for any N -dimensional
closed interval I ⊂ (0, ∞)N , B H,K has a local time L(x, I), x ∈ Rd . Moreover, the local time
admits the following L2 -representation
Z
Z
H,K
L(x, I) = (2π)−d
e−ihy,xi eihy,B (s)i dsdy, x ∈ Rd .
(4.23)
Rd
I
Remark 4.10 Although the existence of local times can also be proved by using the Malliavin
calculus [see Proposition 4.15 below], we prefer to provide a Fourier analytic proof because:
1) we can compare in this way the two methods and 2) the above theorem gives in addition
the representation (4.23).
Proof Without loss of generality, we may assume that I = [ε, 1]N where ε > 0. Let λN be
the Lebesgue measure on I. We denote by µ the image measure of λN under the mapping
t 7→ B H,K (t). Then the Fourier transform of µ is
Z
H,K
µ
b(ξ) = eihξ, B (t)i dt.
(4.24)
I
It follows from Fubini’s theorem and (4.17) that
Z
Z Z Z
³
´
¯
¯2
H,K
H,K
¯
¯
E
µ
b(ξ) dξ =
E eihξ, B (s)−B (t)i dξ dsdt
Rd
IZ IZ Rd
1
=c
£ ¡ H,K
¢2 ¤d/2 dsdt
I I E B
(s) − B1H,K (t)
1
Z Z
1
≤c
£ PN
¤ dsdt.
2Hj Kj d/2
I I
|s
−
t
|
j
j
j=1
(4.25)
The same argument
shows that the last integral is finite
PN in1 Xiao and Zhang (2002, p. 214)
2
d
whenever d < j=1 Hj Kj . Hence, in this case, µ
b ∈ L (R ) a.s. and Theorem 4.9 follows from
the Plancherel theorem.
¤
Remark 4.11 Recently, Ayache, Wu and Xiao (2006) have shown that fractional Brownian
sheets have jointly continuous local times based on the “sectorial local nondeterminism”.
It would be interesting to prove that B H,K is sectorially locally nondeterministic and to
establish joint continuity and sharp H¨older conditions for the local times of B H,K .
Now we consider the Hausdorff and packing dimensions of the image, graph and level set
of B H,K . In order to state our theorems conveniently, we assume
0 < H1 K1 ≤ . . . ≤ HN KN < 1.
(4.26)
We denote packing dimension by dimP ; see Falconer (1990) for its definition and properties. The following theorems can be proved by using Lemma 4.8 and the same arguments as
in Ayache and Xiao (2005, Section 3). We leave the details to the interested reader.
19
Theorem 4.12 With probability 1,
dimH B
H,K
½
¡
¢
¡
¢
N
H,K
N
[0, 1] = dimP B
[0, 1] = min d;
N
X
j=1
1
Hj Kj
¾
(4.27)
and
¡
¢
¡
¢
dimH GrB H,K [0, 1]N = dimP GrB H,K [0, 1]N
( PN
PN
1
1
if
j=1 Hj Kj
Hj Kj ≤ d,
Pk Hk Kk
Pj=1
Pk
=
k−1
1
if
j=1 Hj Kj + N − k + (1 − Hk Kk )d
j=1 Hj Kj ≤ d <
j=1
1
Hj Kj
,
(4.28)
where
P0
1
j=1 Hj Kj
:= 0.
Theorem 4.13 Let Lx = {t ∈ (0, ∞)N : B H,K (t) = x} be the level set of B H,K . The
following statements hold:
P
1
d
(i) If N
j=1 Hj < d, then for every x ∈ R we have Lx = ∅ a.s.
(ii) If
PN
1
j=1 Hj
> d, then for any x ∈ Rd and 0 < ε < 1, with positive probability
¡
¢
¡
¢
dimH Lx ∩ [ε, 1]N = dimP Lx ∩ [ε, 1]N
¾
½X
k
Hk
+ N − k − Hk d, 1 ≤ k ≤ N
= min
Hj
(4.29)
j=1
=
k
X
Hk
j=1
4.4
Hj
+ N − k − Hk d,
if
k−1
k
X
X
1
1
≤d<
.
Hj
Hj
j=1
j=1
A Malliavin calculus approach
Using the Malliavin calculus approach, we can study the local times of more general bifractional Brownian sheets. Consider the (N × d)-matrices
H = (H 1 , . . . , H d )
and
K = (K 1 , . . . , K d ),
and
K i = (Ki,1 , . . . , Ki,N )
where for any i = 1, . . . , d
H i = (Hi,1 , . . . , Hi,N )
with Hi,j ∈ (0, 1) and Ki,j ∈ (0, 1] for every i = 1, . . . , d and j = 1, . . . , N .
20
We will say that the Gaussian field B H,K is an (N, d)-bifractional Brownian sheet with
indices H and K if
³
´
B H,K (t) = B H 1 (t), . . . , B H d (t) , t ∈ [0, ∞)N
and for every i = 1, . . . , d, the process {B H i (t), t ∈ RN
+ } is centered and has covariance
function
N
³
´
Y
RHi,j ,Ki,j (sj , tj ).
E B H i ,K i (t)B H i ,K i (s) = RH i ,K i (s, t) =
j=1
d
H,K is defined as
As in subsection 4.1, the local time L(x, t) ( t ∈ RN
+ and x ∈ R ) of B
the density of the occupation measure µt , defined by
Z
¢
¡
µt (A) =
1lA B H,K (s) ds, A ∈ B(Rd ).
[0,t]
Formally, we can write
Z
L(x, t) =
[0,t]
¡
¢
δx B H,K (s) ds,
where δx denotes the Dirac function and δx (BsH,K ) is therefore a distribution in the Watanabe
sense (see Watanabe (1984)).
We need some notation. For x ∈ R, let pσ (x) be the centered Gaussian kernel with
variance σ > 0. Consider also the Gaussian kernel on Rd given by
pdσ (x)
=
d
Y
pσ (xi ),
x = (x1 , . . . , xd ) ∈ Rd .
i=1
Denote by Hn (x) the n–th Hermite polynomial defined by H0 (x) = 1 and for n ≥ 1,
Hn (x) =
³ x2 ´ dn
³ x2 ´
(−1)n
exp
exp
−
,
n!
2 dxn
2
x ∈ R.
We will make use of the following technical lemma.
Lemma 4.14 For any H ∈ (0, 1) and K ∈ (0, 1], let us define the function
QH,K (z) =
RH,K (1, z)
,
z HK
z ∈ (0, 1]
and QH,K (0) = 0. Then the function QH,K takes values in [0, 1], QH,K (1) = 1 and it is
strictly increasing. Moreover, there exists δ > 0 such that for all z ∈ (1 − δ, 1),
¡
¢
(QH,K (z))n ≤ exp −c(δ, H, K)n(1 − z)2H .
(4.30)
21
Proof
Clearly, the Cauchy-Schwarz inequality implies 0 ≤ QH,K (z) ≤ 1. Let us prove
that the function QH,K is strictly increasing. By computing the derivative Q0H,K (z) and
multiplying this by z HK+1 , we observe that this is equivalent to show
(1 − z)2HK−1 (1 + z) − (1 + z 2H )K−1 (1 − z 2H ) > 0
for all z ∈ (0, 1).
(4.31)
1
If HK ≤
(1 + z 2H )K−1¢ ≤ 1 + z, the left side in (4.31) can be minorized by
¡ 2 , since
2H
K
2HK−1
(1 + z ) (1 − z)
− 1 + z 2H and this is positive since (1 − z)2HK−1 ≥ 1.
If HK > 12 , we note that
(1 − z)2HK−1 (1 + z) + (1 + z 2H )K−1 z 2H ≥ (1 − z)(1 + z) + (1 + z 2H )K−1 z 2
≥ (1 + z 2H )K−1 (1 − z 2 ) + (1 + z 2H )K−1 z 2 ≥ (1 + z 2H )K−1 .
and this implies (4.31). Concerning the inequality (4.30), we note that
QH,K (z)n = exp (n log QH,K (z)) ≥ exp (−n(1 − QH,K (z))) .
Now by Taylor’s formula
(1 + z 2H )K z −HK ≤ 2K + c(H, K, δ)(1 − z)2
and therefore
QH,K (z)
1
(1 − z)2HK
2K
1
≤ 1 + c(H, K, δ)(1 − z)2HK δ 2−2HK − K (1 − z)2HK .
2
≤ 1 + c(H, K, δ)(1 − z)2 −
The conclusion follows as in the proof of Lemma 2 in Eddahbi et al. (2005), since
1 − QH,K (z) ≥
1
(1 − z)2HK (1 − c(H, K, δ))
2K
for any z ∈ (1 − δ, 1) with δ close to zero and with c(H, K, δ) tending to zero as δ → 0.
¤
The following proposition gives a chaotic expansion of the local time of the (N, d)bifractional Brownian sheet. The stochastic integral In (h) appeared below is the multiple
Wiener-Itˆo integral of order n of the function h of nN variables with respect to an (N, 1)
bifractional Brownian motion with parameters H = (H1 , . . . , HN ) and K = (K1 , . . . , KN ) .
Recall that such integrals can be constructed in general on a Gaussian space [see, for example,
Major (1981), or Nualart (1995)]. We will only need the following isometry formula:
N
´
³
Y
¡ H ,K
¢n
⊗m
H,K
n
)
=
R
(t,
s)
1
l
=
R j j (tj , sj ) 1l(n=m)
)I
(1
l
E In (1l⊗n
(n=m)
[0,t] m [0,s]
j=1
for all s, t ∈ RN
+.
22
(4.32)
Proposition 4.15 For any x ∈ Rd and t ∈ (0, ∞)N , the local times L(x, t) admits the
following chaotic expansion
Z
X
L(x, t) =
d p
(xi )
Y
s2H i K
i
sni H i K i
n1 ,...,nd ≥0 [0, t] i=1
Hn i
³ x ´
i
Ini i (1l[0,s] (·)⊗ni ) ds,
H
i
s
(4.33)
Q
Hi,j Ki,j
. The integrals Ini i denotes the multiple Itˆ
where s = s1 · · · sN and sH i K i = N
o
j=1 sj
stochastic integrals with respect to the independent N -parameter bifractional Brownian motion
B H i ,K i .
P
1
∗
∗
Moreover, if N
j=1 H ∗ K ∗ > d, where Hj = max{Hi,j : i = 1, . . . , d} and Kj = max{Ki,j :
j
j
i = 1, . . . , d}, then L(x, t) is a random variable in L2 (Ω).
Proof The chaotic expression (4.33) can be obtained similarly as in Eddahbi et al. (2005)
or Russo and Tudor (2006). It is based on the approximation of the Dirac delta function
by Gaussian kernels with variance converging to zero. Let us evaluate the L2 (Ω) norm of
L(x, t). By the independence of components and the isometry of multiple stochastic integrals,
we obtain
kL(x,
t)k22
=
X
Z
X
Z
du
m≥0 n1 +···+nd =m [0,t]
where
βni (u) =
dv
[0,t]
d
Y
βni (u)βni (v)RH i ,K i (u, v)ni ,
(4.34)
i=1
ps2H i K (xi )
i
sni H i K i
µ
Hni
¶
xi
sH i K i
By Propositions 3 and 6 in Imkeller et al. (1995) [see also Lemma 11 in Eddahbi et al.
(1995)], we have the bound
1
βni (u)βni (v) ≤ c4,9
1
(ni ∨ 1)
8β−1
6
(4.35)
uni H i K i v ni H i K i
for any β ∈ [ 14 , 21 ). Using the inequality (4.35), we derive from (4.34) that kL(x, t)k22 is at
most
c
X
X
m≥0 n1 +···+nd =m
=c
X
µY
d
i=1
m≥0 n1 +···+nd =m
= c4,10 t
2
X
(ni ∨ 1)
µY
d
X
¶Z
1
X
m≥0 n1 +···+nd =m
du
8β−1
6
(ni ∨ 1)
µY
d
i=1
dv
[0,t]
[0,u]
¶Y
N Z
1
i=1
Z
8β−1
6
j=1 0
(ni ∨ 1)
Z
tj
uj duj
¶Y
N Z
1
8β−1
6
j=1 0
23
d Y
N
Y
RHi,j ,Ki,j (uj , vj )ni
(uj vj )ni Hi,j Ki,j
i=1 j=1
d
1µY
0
d
1µY
i=1
¶
QHi,j ,Ki,j (z)
dz
ni
i=1
¶
ni
QHi,j ,Ki,j (z)
dz,
(4.36)
where we used the change of variables uj = uj and vj = zj uj . Using the above lemma and
as in the proof of Lemma 2 in Eddahbi et al. (2005), we can prove the bound
¶
Z 1µY
d
− ∗1 ∗
ni
(4.37)
QHi,j ,Ki,j (z)
dz ≤ c4,11 m 2Hj Kj .
0
i=1
Here c4,11 = c4,11 (H, K) depends on H, K. Finally, (4.37) implies that
kL(x,
t)k22
≤ c4,12
N
X µY
m≥1
≤ c4,13
X
m
m
− 2H ∗1K ∗
j=1
P
− N
j=1
j
¶
µY
d
X
n1 +···+nd =m
1
+d(1− 8β−1
)−1
6
2H ∗ K ∗
j j
i=1
¶
1
j
(ni ∨ 1)
8β−1
6
(4.38)
,
m≥1
where c4,12 and c4,13 depend on H, K and t only. The last series in (4.38) converges if
N
X
j=1
µ
¶
1
8β − 1
>d 1−
.
2Hj∗ Kj∗
6
To conclude, observe that by choosing β close to
1
2,
condition (4.39).
(4.39)
PN
1
j=1 Hj∗ Kj∗
> d implies the required
¤
P
We recall that a random variable F = n In (fn ) belongs to the Watanabe space Dα,2 if
X
kF k2α,2 :=
(1 + m)α kIn (fn )k22 < ∞.
n≥0
Corollary 4.16 For any t ∈ (0, ∞)N and x ∈ Rd , the local time L(x, t) of the (N, d)bifractional
Brownian sheet B H,K belongs to the Watanabe space Dα,2 for every 0 < α <
PN
d
1
j=1 2H ∗ K ∗ − 2 .
j
j
Proof
This is a consequence of the proof of Proposition 4.15. Using the computation
contained there, we obtain for any β ∈ [ 14 , 12 ),
kL(x, t)k2α,2 ≤ c4,14 (H, K, d, t)
X
P
d(1− 8β−1
)−1− N
j=1
6
(1 + m)α m
1
2H ∗ K ∗
j j
m≥1
which is convergent if α <
to
1
2,
PN
1
j=1 2Hj∗ Kj∗
− d(1 −
we get the conclusion.
8β−1
6 )
−1−
PN
1
j=1 2Hj∗ Kj∗ .
Choosing β close
¤
Acknowledgment This work was initiated while both authors were attending the Second Conference on Self-similarity and Applications held during June 20–24, 2005, at INSA
Toulouse, France. We thank the organizers, especially Professor Serge Cohen, for their invitation and hospitality.
24
References
[1] R. Addie, P. Mannersalo and I. Norros (2002), Performance formulae for queues with Gaussian
input. European Trans. Telecommunications 13(3), 183–196.
[2] T. W. Anderson (1955), The integral of a symmetric unimodal function over a symmetric
convex set and some probability inequalities. Proc. Amer. Math. Soc. 6, 170–176.
[3] V. V. Anh, J. M. Angulo and M. D. Ruiz-Medina (1999), Possible long-range dependence in
fractional random fields. J. Statist. Plann. Inference 80, 95–110.
[4] A. Ayache, D. Wu and Y. Xiao (2006), Joint continuity of the local times of fractional Brownian
sheets. Submitted.
[5] A. Ayache and Y. Xiao (2005), Asymptotic properties and Hausdorff dimensions of fractional
Brownian sheets. J. Fourier Anal. Appl. 11, 407–439.
[6] J. M. Aza¨ıs and M. Wschebor (1996), Almost sure oscillation of certain random processes.
Bernoulli 2, 257–270.
[7] A. Benassi, P. Bertrand et J. Istas (2000), Identification of the Hurst exponent of a Step
Multifractional Brownian motion. Statistical Inference for Stochastic Processes, 13, 101–111.
[8] D. A. Benson, M. M. Meerschaert and B. Baeumer (2006), Aquifer operator-scaling and the
effect on solute mixing and dispersion. Water Resour. Res. 42, W01415.
[9] S. M. Berman (1972), Gaussian sample function: uniform dimension and H¨older conditions
nowhere. Nagoya Math. J. 46, 63–86.
[10] S. M. Berman (1973), Local nondeterminism and local times of Gaussian processes. Indiana
Univ. Math. J. 23, 69–94.
[11] N. H. Bingham, C. M. Goldie and J. L. Teugels (1987), Regular Variation. Cambridge University Press, Cambridge.
[12] A. Bonami and A. Estrade (2003), Anisotropic analysis of some Gaussian models. J. Fourier
Anal. Appl. 9, 215–236.
[13] P. Cheridito (2004), Gaussian moving averages, semimartingales and option pricing. Stochastic
Process. Appl. 109, 47–68.
[14] P. Cheridito, H. Kawaguchi and M. Maejima (2003), Fractional Ornstein-Uhlenbeck processes.
Electronic Journal of Probability, 8, paper 3, pp. 1–14.
[15] I. P. Cornfeld, S. V. Fomin and Ya. G. Sinai (1982), Ergodic Theory. Springer, New York.
[16] J. Cuzick and J. DuPreez (1982), Joint continuity of Gaussian local times. Ann. Probab. 10,
810–817.
[17] M. Eddahbi, R. Lacayo, J. L. Sol´e, C. A. Tudor, J. Vives (2005), Regularity of the local time
for the d–dimensional fractional Brownian motion with N -parameters. Stoch. Anal. Appl., 23,
no. 2, 383–400.
[18] K. J. Falconer (1990), Fractal Geometry – Mathematical Foundations and Applications. Wiley
& Sons, Chichester.
[19] D. Geman and J. Horowitz (1980), Occupation densities. Ann. Probab. 8, 1–67.
[20] C. Houdr´e and J. Villa (2003), An example of infinite dimensional quasi-helix. Stochastic models
(Mexico City, 2002), pp.195–201, Contemp. Math., 336, Amer. Math. Soc., Providence, RI.
25
[21] P. Imkeller, V. Perez–Abreu and J. Vives (1995), Chaos expansion of double intersection local
time of Brownian motion in Rd and renormalization. Stochastic Process. Appl. 56, 1–34.
[22] J.-P. Kahane (1985), Some Random Series of Functions. 2nd edition, Cambridge University
Press, Cambridge.
[23] Y. Kasahara, N. Kˆono and T. Ogawa (1999), On tail probability of local times of Gaussian
processes. Stochastic Process. Appl. 82, 15–21.
[24] J. Kuelbs, W. V. Li and Q.-M. Shao (1995), Small ball probabilities for Gaussian processes
with stationary increments under H¨older norms. J. Theoret. Probab. 8, 361–386.
[25] J. Lamperti (1962), Semi-stable stochastic processes. Trans. Amer. Math. Soc. 104, 64–78.
[26] W. V. Li and Q.-M. Shao (2001), Gaussian processes: inequalities, small ball probabilities and
applications. In Stochastic Processes: Theory and Methods. Handbook of Statistics, 19, (C. R.
Rao and D. Shanbhag, editors), pp. 533–597, North-Holland.
[27] L, Lo´eve (1977), Probability Theory I. Springer, New York.
[28] P. Major (1981), Multiple Wiene-Itˆ
o Integrals. Lecture Notes in Math. 849, Sringer-Verlag,
Berlin.
[29] P. Mannersalo and I. Norros (2002), A most probable path approach to queueing systems with
general Gaussian input. Comp. Networks 40 (3), 399–412.
[30] G. Maruyama (1949), The harmonic analysis of stationary stochastic processes. Mem. Fac.
Sci. Kyushu Univ. A 4, 45–106.
[31] D. Monrad and L. D. Pitt (1987), Local nondeterminism and Hausdorff dimension. In: Progress
in Probability and Statistics. Seminar on Stochastic Processes 1986, (E, Cinlar, K. L. Chung,
R. K. Getoor, Editors), pp.163–189, Birkhauser, Boston.
[32] D. Monrad and H. Rootz´en (1995), Small values of Gaussian processes and functional laws of
the iterated logarithm. Probab. Theory Relat. Fields 101, 173–192.
[33] D. Nualart (1995), Malliavin Calculus and Related Topics. Springer, New York.
[34] E. J. G. Pitman (1968), On the behavior of the characteristic function of a probability sidtribution in the neighbourhood of the origin. J. Australian Math. Soc. Series A 8, 422–443.
[35] L. D. Pitt (1978), Local times for Gaussian vector fields. Indiana Univ. Math. J. 27, 309–330.
[36] F. Russo and C. A. Tudor (2006), On the bifractional Brownian motion. Stochastic Process.
Appl. 5, 830–856.
[37] G. Samorodnitsky and M. S. Taqqu (1994), Stable non-Gaussian Random Processes: Stochastic
models with infinite variance. Chapman & Hall, New York.
[38] N.-R. Shieh and Y. Xiao (2005), Images of Gaussian random fields: Salem sets and interior
points. Studia Math. to appear.
[39] W. Stolz (1996), Some small ball probabilities for Gaussian processes under nonuniform norms.
J. Theoret. Probab. 9, 613–630.
[40] K. Takashima (1989), Sample path properties of ergodic self-similar processes. Osaka Math. J.
26, 159–189.
[41] M. Talagrand (1995), Hausdorff measure of trajectories of multiparameter fractional Brownian
motion. Ann. Probab. 23, 767–775.
26
[42] S. Watanabe (1984), Lectures on Stochastic Differential Equations and Malliavin Calculus.
Springer, Berlin.
[43] M. Wschebor (1992), Sur les accroissements du processus de Wiener. C. R. Acad. Sci. Paris
315, 1293–1296.
[44] Y. Xiao (1997), H¨older conditions for the local times and the Hausdorff measure of the level
sets of Gaussian random fields. Probab. Theory Relat. Fields 109, 129–157.
[45] Y. Xiao (2005), Strong local nondeterminism and the sample path properties of Gaussian
random fields. Submitted.
[46] Y. Xiao and T. Zhang (2002), Local times of fractional Brownian sheets. Probab. Theory Relat.
Fields 124, 204–226.
27